【作者】 刘敏；

【导师】 王忠杰；

【作者基本信息】 哈尔滨工业大学 ， 软件工程， 2022， 硕士

【摘要】 随着互联网的普及以及智能设备的迅速发展,虚拟个人助手(Virtual Personal Assistant,VPA),也称软服务机器人,逐渐进入人们的生活的方方面面,提供便利。同时,会话推荐系统,也即基于对话的服务推荐系统,以VPA作为载体,也逐渐成为研究热点。会话推荐系统以其可及时获得用户反馈的特点,能够更加精准的命中用户的需求。同时也带来以下几个挑战:如何捕捉用户偏好并对其建模?面对海量服务如何高效推荐,减少响应时间?如何制定合适的对话策略以提高推荐效率?本文将主要针对这三个问题进行研究。(1)本文基于传统用户画像模型,提出了基于人机对话的精细化用户画像(Conversational Bot based Elaborate User Portrait,CBEUP),以更加精准的表达用户的偏好。为了在尽量少的对话轮次内给用户推荐心仪的服务,可以利用用户画像对用户长期的偏好进行建模。传统的标签化用户画像难以对数值型的偏好进行精确的表达,并且难以记录用户偏好的更迭。因此,该模型记录了用户偏好的迭代历史、不同偏好的数值化约束、偏好的提及频度,并且通过语义连接了多个领域的偏好。(2)本文提出了服务特征依赖模型,以及相应的服务预排序算法,为服务推荐提供支持。为了缩减候选服务规模,本文考虑到多领域服务之间所存在的客观依赖,例如空间依赖,我们认为用户在选择了某一领域的服务后,在另一领域更倾向于选择与其依赖程度密切的服务。本文通过对不同领域间的服务之间的依赖紧密度进行度量,用于缩减对话过程中候选服务的规模,以提高服务推荐的效率。(3)本文提出了领域间通过回溯来确定需要进行推荐的领域,领域内基于反馈迭代推荐的服务推荐策略,并最终构造了相应的服务推荐VPA(service recommendation VPA,sr VPA)。服务推荐策略用于引导用户针对不同领域间的服务进行对话,解决了由于服务依赖模型的引入,当用户选择某个领域的服务时,其它领域可能缺少与已选择的服务依赖紧密的候选服务的问题。(4)本文提供了面向服务推荐的软服务机器人sr VPA的设计与实现细节,包括架构、功能设计,以及系统实现的类图以及业务逻辑设计。通过不同的对话案例验证了系统各项功能的实现稳定性。更多还原

【Abstract】 With the popularization of the Internet and the rapid development of smart devices,virtual personal assistants(VPAs),also known as soft service robots,have gradually entered all aspects of people’s lives to provide convenience.At the same time,the conversational recommender system(CRS),that is,the dialogue-based service recommendation system,with VPA as the carrier,has gradually become a research hotspot.CRS can more accurately target users’ needs due to the fact that they can obtain user feedback in time.It also brings the following challenges: How to capture and model user preferences? How to efficiently recommend in the face of massive services to reduce response time? How to formulate an appropriate dialogue strategy to improve recommendation efficiency? This paper will mainly focus on these three questions.(1)This paper proposes the Conversational Bot based Elaborate User Portrait(CBEUP)based on human-machine dialogue to more accurately express user preferences.In order to recommend services of interest to users in fewer dialogue rounds,user portraits can be used to model users’ long-term preferences.It’s difficult for the traditional labelled user portrait to accurately express numerical preferences and record the changes of user preferences.Therefore,CBEUP records the iterative update history,the numerical constraints and the mention frequency of preferences,and semantically connects multi-domain preferences.(2)This paper proposes the service feature dependency model and a corresponding service pre-sorting algorithm to support service recommendation.In order to reduce the scale of candidate services,this paper considers the objective dependencies between multidomain services,such as spatial dependencies,and believe that after users choose a service in a certain domain,they are more inclined to choose a service that is closely dependent on it in another domain.Therefore,this paper measure the closeness of dependency between services in different domains to reduce the scale of candidate services in the dialogue process and improve the efficiency of service recommendation.(3)This paper proposes a service recommendation policy that uses backtracking method to determine the domain that needs to be recommended,and iterative recommends based on user feedback within the domain.Finally,this paper builds a service recommendation VPA(sr VPA).The service recommendation policy is used to guide users to conduct dialogues on services among different domains,which solves the problem that due to the introduction of the service feature dependency model,when a user selects a service in a certain domain,there is a possibility that other domains lack candidate services that are closely dependent on the selected service.(4)This paper provides the design and implementation details of the service recommendation-oriented soft service robot(sr VPA),including the architecture and function design,as well as the specific implementation class diagram and business logic design.The realization stability of each function of the system is verified through different dialogue cases.更多还原